On Wed, 2017-03-22 at 15:47 -0700, Francisco Jerez wrote:
> Samuel Iglesias Gonsálvez <sigles...@igalia.com> writes:
>
> > Generalize it to lower any unsupported narrower conversion.
> >
> > v2 (Curro):
> > - Add supports_type_conversion()
> > - Reuse existing intruction instead of cloning it.
> > - Generalize d2x to narrower and equal size conversions.
> >
> > Signed-off-by: Samuel Iglesias Gonsálvez <sigles...@igalia.com>
> > ---
> > src/intel/compiler/brw_fs.cpp | 11 ++--
> > src/intel/compiler/brw_fs_lower_d2x.cpp | 97
> > ++++++++++++++++++++++++---------
> > 2 files changed, 77 insertions(+), 31 deletions(-)
> >
> > diff --git a/src/intel/compiler/brw_fs.cpp
> > b/src/intel/compiler/brw_fs.cpp
> > index 6da23b17107..8612a83805d 100644
> > --- a/src/intel/compiler/brw_fs.cpp
> > +++ b/src/intel/compiler/brw_fs.cpp
> > @@ -5694,11 +5694,6 @@ fs_visitor::optimize()
> > OPT(dead_code_eliminate);
> > }
> >
> > - if (OPT(lower_d2x)) {
> > - OPT(opt_copy_propagation);
> > - OPT(dead_code_eliminate);
> > - }
> > -
> > OPT(lower_simd_width);
> >
> > /* After SIMD lowering just in case we had to unroll the EOT
> > send. */
> > @@ -5745,6 +5740,12 @@ fs_visitor::optimize()
> > OPT(dead_code_eliminate);
> > }
> >
> > + if (OPT(lower_d2x)) {
> > + OPT(opt_copy_propagation);
> > + OPT(dead_code_eliminate);
> > + OPT(lower_simd_width);
> > + }
> > +
> > lower_uniform_pull_constant_loads();
> >
> > validate();
> > diff --git a/src/intel/compiler/brw_fs_lower_d2x.cpp
> > b/src/intel/compiler/brw_fs_lower_d2x.cpp
> > index a2db1154615..fa25d313dcd 100644
> > --- a/src/intel/compiler/brw_fs_lower_d2x.cpp
> > +++ b/src/intel/compiler/brw_fs_lower_d2x.cpp
> > @@ -27,47 +27,92 @@
> >
> > using namespace brw;
> >
> > +static bool
> > +supports_type_conversion(fs_inst *inst) {
>
> Pointer should be marked const.
>
> > + switch(inst->opcode) {
> > + case BRW_OPCODE_MOV:
> > + case SHADER_OPCODE_MOV_INDIRECT:
> > + return true;
> > + case BRW_OPCODE_SEL:
> > + return false;
>
> I suggest you return 'inst->dst.type == get_exec_type_size(inst)'
> here
> in order to simplify the logic below and restructure things in a way
> that allows you to make opcode-specific decisions here based on the
> actual execution and destination types.
I guess you meant:
'type_sz(inst->dst.type) == get_exec_type_size(inst)'
This would make it simpler but we then allow f2d conversion which is
not allowed for SEL.
>
> > + default:
> > + /* FIXME: We assume the opcodes don't explicitly mentioned
> > + * before just work fine with arbitrary conversions.
> > + */
> > + return true;
> > + }
> > +}
> > +
> > bool
> > fs_visitor::lower_d2x()
> > {
> > bool progress = false;
> >
> > foreach_block_and_inst_safe(block, fs_inst, inst, cfg) {
> > - if (inst->opcode != BRW_OPCODE_MOV)
> > - continue;
> > + bool inst_support_conversion =
> > supports_type_conversion(inst);
> > + bool supported_conversion =
> > + inst_support_conversion &&
> > + (get_exec_type_size(inst) != 8 ||
> > + type_sz(inst->dst.type) > 4 ||
> > + type_sz(inst->dst.type) >= get_exec_type_size(inst));
> >
> > - if (inst->dst.type != BRW_REGISTER_TYPE_F &&
> > - inst->dst.type != BRW_REGISTER_TYPE_D &&
> > - inst->dst.type != BRW_REGISTER_TYPE_UD)
> > + /* If the conversion is supported or there is no conversion
> > then
> > + * do nothing.
> > + */
> > + if (supported_conversion ||
> > + (!inst_support_conversion && inst->dst.type == inst-
> > >src[0].type) ||
>
> You should be using the execution type here (and below where you
> allocate the temp0 and temp1 vgrfs) instead of assuming it matches
> the
> type of the first source. Also I'm not 100% certain that the
> combination of this conditional continue, the one below, the if/else
> branches below and the supported_conversion definitions above catch
> the
> exact set of cases where you need to apply lowering. This would be a
> lot easier to follow (most of the above goes away) if you did
> something
> along the lines of:
>
OK, thanks.
> > if (supports_type_conversion(inst)) {
> > if (get_exec_type_size(inst) == 8 && type_sz(inst->dst.type) <
> > 8) {
> > // Apply strided move workaround...
> > progress = true;
> > }
> > } else {
> > // Apply general conversion workaround...
> > progress = true;
> > }
> > + inst->dst.file == BAD_FILE || inst->src[0].file ==
> > BAD_FILE)
>
> I don't think the 'inst->dst.file == BAD_FILE || inst->src[0].file ==
> BAD_FILE' terms of this expression are correct or useful, the
> 'inst->src[0].file' term because there might be a destination region
> conversion regardless of what the file of the first source is, the
> 'inst->dst.file == BAD_FILE' term because you could just rely on DCE
> instead...
>
OK
> > continue;
> >
> > - if (inst->src[0].type != BRW_REGISTER_TYPE_DF &&
> > - inst->src[0].type != BRW_REGISTER_TYPE_UQ &&
> > - inst->src[0].type != BRW_REGISTER_TYPE_Q)
> > - continue;
> > + /* This pass only supports conversion to narrower or equal
> > size types. */
> > + if (get_exec_type_size(inst) < type_sz(inst->dst.type))
> > + continue;
> >
>
> I think you need to apply this condition to the if branch below that
> deals with DF to F conversions *only*, the else branch you probably
> need
> to apply regardless of whether the destination type is larger or
> smaller
> than the execution type.
>
OK
> > - assert(inst->dst.file == VGRF);
> > assert(inst->saturate == false);
>
> I don't see why it's safe to rely on saturate being false here, in
> any
> case it would be straightforward to handle it by setting the flag on
> the
> emitted MOV instruction and clearing it on the original instruction.
>
OK
> > - fs_reg dst = inst->dst;
> >
> > const fs_builder ibld(this, block, inst);
> > + fs_reg dst = inst->dst;
> >
> > - /* From the Broadwell PRM, 3D Media GPGPU, "Double Precision
> > Float to
> > - * Single Precision Float":
> > - *
> > - * The upper Dword of every Qword will be written with
> > undefined
> > - * value when converting DF to F.
> > - *
> > - * So we need to allocate a temporary that's two registers,
> > and then do
> > - * a strided MOV to get the lower DWord of every Qword that
> > has the
> > - * result.
> > - */
> > - fs_reg temp = ibld.vgrf(inst->src[0].type, 1);
> > - fs_reg strided_temp = subscript(temp, inst->dst.type, 0);
> > - ibld.MOV(strided_temp, inst->src[0]);
> > - ibld.MOV(dst, strided_temp);
> > + if (inst_support_conversion && !supported_conversion) {
> > + /* From the Broadwell PRM, 3D Media GPGPU, "Double
> > Precision Float to
> > + * Single Precision Float":
> > + *
> > + * The upper Dword of every Qword will be written with
> > undefined
> > + * value when converting DF to F.
> > + *
> > + * So we need to allocate a temporary that's two
> > registers, and then do
> > + * a strided MOV to get the lower DWord of every Qword
> > that has the
> > + * result.
> > + */
> > + fs_reg temp = ibld.vgrf(inst->src[0].type, 1);
>
> The '1' argument is redundant in all fs_builder::vgrf() calls here
> and
> below.
>
OK
> > + fs_reg strided_temp = subscript(temp, dst.type, 0);
> > +
> > + /* We clone the original instruction as we are going to
> > modify it
> > + * and emit another one after it.
> > + */
>
> Comment seems stale.
>
Right.
> > + inst->dst = strided_temp;
> > + /* As it is an strided destination, we write n-times more
> > being n the
> > + * size difference between source and destination types.
> > Update
>
> I guess by difference you meant ratio here and in the copy-pasted
> comment below (the comments also have a few minor spelling mistakes
> but
> that's unlikely to cause as much confusion).
>
Right.
> > + * size_written with the new destination.
> > + */
> > + inst->size_written = inst->dst.component_size(inst-
> > >exec_size);
>
> I think I'd add an assert(inst->size_written ==
> inst->dst.component_size(inst->exec_size)) right before you assign
> 'inst->dst' to point at the temporary in order to make sure you
> aren't
> accidentally miscompiling instructions that write multiple
> components.
>
OK
> > + ibld.at(block, inst->next).MOV(dst, strided_temp);
> > + } else {
> > + fs_reg temp0 = ibld.vgrf(inst->src[0].type, 1);
> > + fs_reg temp1 = ibld.vgrf(inst->src[0].type, 1);
> > + fs_reg strided_temp1 = subscript(temp1, dst.type, 0);
> > +
> > + inst->dst = temp0;
> > + /* As it is an strided destination, we write n-times more
> > being n the
> > + * size difference between source and destination types.
> > Update
> > + * size_written with the new destination.
> > + */
> > + inst->size_written = inst->dst.component_size(inst-
> > >exec_size);
> >
> > - inst->remove(block);
> > + /* Now, do the conversion to original destination's type.
> > */
> > + fs_inst *mov = ibld.at(block, inst-
> > >next).MOV(strided_temp1, temp0);
> > + ibld.at(block, mov->next).MOV(dst, strided_temp1);
>
> Isn't this MOV and the strided_temp1 handling redundant? I'd expect
> it
> to be handled automatically during the next loop iteration of this
> pass
> (though you may have to switch back to foreach_block_and_inst()
> instead
> of foreach_block_and_inst_safe()), which would also result in better
> code because the extra move would only be emitted if this is an
> actual
> DF->F conversion.
>
Good catch!
Thanks for all the suggestions, I am going to do them.
Sam
> > + }
> > progress = true;
> > }
> >
> > --
> > 2.11.0
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev@lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
signature.asc
Description: This is a digitally signed message part
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
